Urban style and appearance quality evaluation method and system fusing street scene image multi-scale features

Abstract

The invention provides a city style and appearance quality evaluation method and system fusing street scene image multi-scale features, and relates to the technical field of geographic images, and the method comprises the steps: responding to an evaluation request, obtaining street scene images and geographic position information of each sampling point in a to-be-evaluated region, carrying out the preprocessing of the street scene images, and obtaining a city style and appearance quality evaluation result; and inputting the preprocessed streetscape image into a pre-trained style quality evaluation model to obtain a style quality score value of each sampling point, the style quality evaluation model being used for extracting visual features of different scales from the streetscape image for fusion, and outputting the style quality score value based on the fused visual features, and based on the geographical location information of each sampling point and the corresponding style quality score value, obtaining an urban style quality evaluation result of the to-be-evaluated region. The technical problem that the city style and appearance quality evaluation precision is insufficient due to the fact that the city style and appearance quality cannot be accurately and comprehensively evaluated in the prior art is solved, and the effect of improving the city style and appearance quality evaluation precision is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geographic images, in particular to a cityscape quality evaluation method and system fusing multi-scale features of street view images. BACKGROUND

[0002] Cityscape is a concentrated embodiment of city culture, history and spatial form, and is a key dimension for measuring the quality of urban living environment and the level of sustainable development. Scientific and accurate evaluation of cityscape quality has important theoretical and practical significance for guiding urban planning and design, optimizing public space and improving residents' happiness. However, traditional evaluation methods based on expert field research and questionnaire survey have inherent limitations such as strong subjectivity, limited coverage, high cost and low efficiency, which are difficult to meet the needs of modern city large-scale, high-frequency and fine management.

[0003] With the rapid development of geographic information technology and computer vision, it has become a research hotspot to use street view images, an easily accessible and fully covered data source, for quantitative analysis of urban environment. Existing technologies usually use a full convolutional network with an encoder-decoder structure to identify and count elements such as vegetation, buildings, sky and roads in street view images, and evaluate specific dimensions based on a single physical indicator.

[0004] However, the above-mentioned existing technologies have the following core defects when applied to cityscape quality evaluation: First, since the model training target is limited to element identification or binary classification, the evaluation system naturally revolves around a single physical indicator or a binary category, which cannot accommodate multi-dimensional elements such as architectural harmony and spatial permeability, resulting in a one-sided evaluation result that cannot truly reflect the comprehensive perception of people in the street space. Second, the model relied on lacks an explicit multi-scale feature fusion mechanism, and different levels of features cannot effectively interact, resulting in insufficient collaborative analysis of macro patterns and micro details. Finally, the technical path is essentially to first identify physical elements and then perform arithmetic calculations on physical indicators, and the output objective indicators or discrete categories cannot accurately reflect the continuous and delicate level of landscape perception.

[0005] In summary, due to the inherent defects of existing technologies such as single evaluation dimension, shallow feature utilization and the gap between indicators and perception, only primary element evaluation can be achieved, which cannot accurately and comprehensively evaluate cityscape quality, resulting in insufficient cityscape quality evaluation accuracy. SUMMARY

[0006] In view of the above deficiencies in the prior art, the present application aims to provide a cityscape quality evaluation method fusing multi-scale features of street view images, which has the characteristics of improving the accuracy of cityscape quality evaluation.

[0007] The above invention purpose of the present application is achieved by the following technical solutions: A cityscape quality evaluation method fusing multi-scale features of street view images, comprising: In response to an evaluation request, acquiring street view images and geographic position information of each sampling point in a region to be evaluated; Preprocessing the street view images to obtain preprocessed street view images; Inputting the preprocessed street view images into a pre-trained cityscape quality evaluation model to obtain cityscape quality score values of the sampling points, wherein the pre-trained cityscape quality evaluation model is used to fuse visual features of different scales extracted from the preprocessed street view images, and output the cityscape quality score values based on the fused visual features; and based on the geographic position information of the sampling points and the corresponding cityscape quality score values, obtaining a cityscape quality evaluation result of the region to be evaluated.

[0008] By using the above technical solutions, street view images with geographic position information are acquired as multi-dimensional perception data basis, and standardized preprocessing is used to ensure the robustness of model input, the images are input into the evaluation model to obtain continuous quality scores reflecting human comprehensive perception, and finally the scores of discrete sampling points are combined with geographic positions to generate a globally continuous evaluation result through spatial analysis, realizing the full-process automation from massive image data to comprehensive perception quantization and spatial visualization analysis, so that the cityscape quality evaluation can be large-scale covered in breadth, finely analyzed in depth, and more in line with human experience in validity, and the cityscape quality evaluation precision is improved.

[0009] Preferably, the acquisition of the street view images and the geographic position information of each sampling point in the region to be evaluated comprises: Based on a preset road network sampling density, sampling points are arranged on a road network of the region to be evaluated; According to the arranged sampling points, longitude and latitude coordinates of the sampling points are determined as the geographic position information; A street view map service interface is called to acquire street view images of the sampling points based on the geographic position information of the sampling points and a preset collection direction.

[0010] By using the above technical solutions, sampling points are arranged at a preset density according to the road network structure, ensuring systematic coverage of the city passable area, and through calling the map service interface and combining the preset collection direction, multi-view street view images are batch-acquired, and high-quality image data capable of reflecting street facades and spatial environment in all directions are constructed.

[0011] Preferably, the preprocessing of the street view images to obtain preprocessed street view images comprises: determine whether the street view image meets a preset quality standard, and if not, perform rejection processing to obtain an effective street view image; and perform size scaling processing on the effective street view image according to a preset size to obtain an intermediate street view image. perform pixel value standardization processing on the intermediate street view image to obtain the preprocessed street view image.

[0012] By using the above technical solutions, low-quality images are automatically screened and rejected, ensuring the clarity and reliability of the input data and avoiding noise interference. By uniformly scaling the size, spatial standardization is achieved, ensuring consistency in processing and improving computational efficiency. By standardizing the pixel values, the data is adapted to the model training range, accelerating convergence and enhancing generalization ability, thereby constructing high-quality, standardized input data.

[0013] Preferably, the pre-trained style quality evaluation model includes an encoder, a multi-scale feature fusion module, and a regressor. The preprocessed street view image is input into the pre-trained style quality evaluation model to obtain the style quality score value of each sampling point, which includes: The preprocessed street view image is input into the encoder, and the encoder performs multi-level feature extraction on the preprocessed street view image to obtain visual feature maps of different scales. The visual feature maps of different scales are input into the multi-scale feature fusion module, and the multi-scale feature fusion module performs fusion processing on the visual feature maps to obtain a target visual feature map. The target visual feature map is input into the regressor, and the regressor performs mapping and regression calculation on the target visual feature map to obtain the style quality score value.

[0014] By using the above technical solutions, the encoder is used to extract deep convolutional features from the street view image, obtaining visual feature maps of different scales containing micro details to macro semantics. The multi-scale feature fusion module deeply fuses and complements the rich semantic information of deep features with the accurate spatial details of shallow features, generating a comprehensive target feature map with high semantic expression and high spatial resolution. Finally, the regressor performs global pooling and non-linear mapping on the fused comprehensive features, directly converting them into a continuous style quality score value. This realizes end-to-end modeling from raw image pixels to high-level perception scores, enabling the model to output comprehensive quality evaluation results highly consistent with human subjective judgments based on comprehensive visual representations that fuse multi-scale information.

[0015] Preferably, the visual feature maps of different scales include shallow feature maps and deep feature maps. The different scale visual feature maps are input into the multi-scale feature fusion module, the visual feature maps are processed by the multi-scale feature fusion module, and a target visual feature map is obtained. The channel number adjustment is performed on the shallow feature map and the deep feature map, and the adjusted shallow feature map and deep feature map are obtained. The adjusted deep feature map is sequentially up-sampled to the size of the adjacent shallow feature map, and the feature addition is performed on the corresponding shallow feature map, and a multi-level fusion feature map is obtained. The multi-level fusion feature map is input into a preset convolutional neural network block for feature enhancement processing, and the target visual feature map is obtained.

[0016] By adopting the above technical solution, the feature dimension is unified by channel adjustment, the deep semantic information and the shallow spatial details are gradually fused by the cascade up-sampling and feature addition operation, the multi-level complementary fusion feature representation is constructed, and the high-quality target feature map that can cooperatively reflect the macro, meso and micro features of the cityscape is finally generated.

[0017] Preferably, the target visual feature map is input into the regressor, the target visual feature map is mapped and calculated by the regressor, and the cityscape quality score value is obtained. The global average pooling processing is performed on the target visual feature map, and a global feature vector is obtained. The global feature vector is input into a multi-layer fully connected network, the global feature vector is dimensionally reduced and nonlinearly transformed by the multi-layer fully connected network, and the cityscape quality score value is obtained.

[0018] By adopting the above technical solution, the target visual feature map with spatial dimension is condensed into a compact global feature vector by global average pooling processing. The vector aggregates the visual information of all positions in the global map, effectively maintains the global consistency of the feature, reduces the parameter quantity, enhances the robustness of the model, and gradually reduces the dimension and nonlinearly transforms the global feature vector by the multi-layer fully connected network. The key perception factors highly related to the cityscape quality score are gradually abstracted and refined, and finally the high-dimensional fusion visual feature is mapped into a continuous and single quality score value.

[0019] Preferably, the cityscape quality evaluation result of the to-be-evaluated region is obtained based on the geographical position information of each sampling point and the corresponding cityscape quality score value. Based on the geographical position information of the sampling points and the corresponding landscape quality score values, a landscape quality continuous distribution map covering the to-be-evaluated region is generated by a spatial interpolation algorithm, as the urban landscape quality evaluation result of the to-be-evaluated region.

[0020] By adopting the technical solution, the quantitative landscape quality score obtained on the discrete sampling points is combined with the accurate geographical position coordinates, and a continuous and smooth landscape quality distribution map covering the entire to-be-evaluated region is derived and generated by using a spatial interpolation algorithm, so that the visual conversion of the evaluation result from discrete point data to continuous spatial field is realized, the high-low difference and spatial distribution pattern of the landscape quality at different positions in the region can be intuitively and panoramically displayed, and a direct visual basis is provided for identifying the landscape advantage area, short board area and gradient change rule, so that the spatial readability and decision support value of the evaluation result are greatly enhanced.

[0021] The second aspect of the present application is to provide a kind of urban landscape quality evaluation system of fusion street view image multi-scale feature, with the characteristics of improving urban landscape quality evaluation precision.

[0022] The second aspect of the above application of the present application is realized by the following technical solution: A kind of urban landscape quality evaluation system of fusion street view image multi-scale feature, comprising: Data acquisition module, for responding to evaluation request, the street view image and geographical position information of each sampling point in the to-be-evaluated region are acquired;Image pre-processing module is used for pre-processing the street view image, and the pre-processed street view image is obtained; Quality score module, for inputting the pre-processed street view image into pre-trained landscape quality evaluation model, and obtaining the landscape quality score value of each sampling point, wherein the pre-trained landscape quality evaluation model is used to extract different scale visual features from the pre-processed street view image for fusion, and outputs the landscape quality score value based on the fused visual features; Evaluation result generation module, for obtaining the urban landscape quality evaluation result of the to-be-evaluated region based on the geographical position information of the sampling points and the corresponding landscape quality score values.

[0023] By adopting the above technical solution, each module cooperates in turn, realizes the full-process automation from mass image data to comprehensive perception quantization and then to spatial visualization analysis, so that the urban landscape quality evaluation can be large-scale covered in breadth, finely analyzed in depth, and more in line with human experience in validity, and the urban landscape quality evaluation precision is improved.

[0024] The third aspect of the present application is to provide an electronic device with improved urban landscape quality evaluation precision.

[0025] The third application purpose of the present application is achieved by the following technical solution: An electronic device comprising a memory and a processor, the memory storing a computer program capable of being loaded by the processor and executing the cityscape quality evaluation method of fusing multi-scale features of street view images according to any one of the preceding application purposes.

[0026] The fourth application purpose of the present application is to provide a computer storage medium capable of storing a corresponding program and having the characteristics of facilitating the improvement of the precision of cityscape quality evaluation.

[0027] The fourth application purpose of the present application is achieved by the following technical solution: A computer readable storage medium storing a computer program capable of being loaded by a processor and executing the cityscape quality evaluation method of fusing multi-scale features of street view images according to any one of the preceding application purposes.

[0028] In summary, the present application has at least one of the following beneficial technical effects: 1. The present application realizes the full automation of the evaluation process and the visualization of the results, improves the coverage, depth of analysis and result validity of cityscape evaluation, and thus effectively improves the precision and efficiency of cityscape quality evaluation; 2. The present application extracts multi-scale visual features through an encoder, realizes the deep fusion of deep semantics and shallow details through a multi-scale feature fusion module, generates a comprehensive feature map with high semantic information and high spatial precision, and then directly maps the continuous quality score through a regressor, thereby improving the accuracy and consistency of the evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a step flowchart of the existing technology using a pre-trained semantic segmentation network to score the greening quality.

[0030] Figure 2 is a step flowchart of a cityscape quality evaluation method of fusing multi-scale features of street view images provided by the first embodiment of the present application.

[0031] Figure 3 is a step flowchart of another cityscape quality evaluation method of fusing multi-scale features of street view images provided by the second embodiment of the present application.

[0032] Figure 4 is a structural block diagram of a cityscape quality evaluation system of fusing multi-scale features of street view images provided by the third embodiment of the present application. DETAILED DESCRIPTION

[0033] Please refer to Figure 1The technical solution of the prior art is to first input a single street view image, and then perform forward propagation through a pre-trained semantic segmentation network (such as DeepLabV3+). The network consists of an encoder and a decoder: the encoder, usually composed of a CNN (such as ResNet) backbone, is responsible for feature extraction. Its deep output high semantic, low resolution feature map (such as C5) is used to identify the category of vegetation, buildings and other targets.

[0034] The decoder: the features output by the encoder are upsampled and connected by jump, gradually recovering the spatial resolution, and finally outputting a pixel-level classification result with the same size as the input image, i.e. a semantic segmentation map.

[0035] Post-processing module: according to the segmentation result, the proportion of vegetation category pixels in the total pixels is calculated, and the green view rate index is calculated, which is used as the final evaluation basis for the green quality of urban streets.

[0036] The signal transmission relationship between the components is: street view image → encoder → decoder → semantic segmentation map → green view rate calculation → green quality score.

[0037] However, when used for the more comprehensive evaluation task of urban landscape quality, the above technical solution has the following specific shortcomings: 1. The evaluation dimension is single, resulting in a one-sided evaluation result that cannot reflect the comprehensive quality of urban landscape.

[0038] Specific shortcomings: the existing technical solution only outputs a single physical indicator, green view rate, and its evaluation system is completely focused on greenery. However, the quality of urban landscape is an organic whole composed of multiple elements such as greenery, architecture, street furniture, and spatial form. Evaluating only greenery and ignoring other elements and their mutual coordination makes the system give a one-sided high score that is contrary to human's real perception when faced with a street with a dilapidated building but good greenery.

[0039] Reason for the defect: the technical target itself is limited to green evaluation. Therefore, the training target of the model structure (semantic segmentation network) and the final post-processing logic only serve the extraction and statistics of information about vegetation, and are not designed to perceive and integrate information about other landscape elements. This is an inherent limitation of the original design of the technical solution.

[0040] 2. The model's feature utilization is not sufficient, and there is a lack of multi-scale information fusion, resulting in insufficient ability to analyze the spatial structure and details of the landscape.

[0041] Specific defects: the encoder-decoder structure relied on by the prior art fuses part of the features of different levels through skip connection, but this fusion is passive and shallow, and its main purpose is to optimize the accuracy of the segmentation boundary, rather than to understand the multi-scale landscape semantics. Therefore, the model is difficult to accurately capture the macro urban skyline, the meso building group layout relationship and the micro building facade material or green level, resulting in that the segmentation result output by the model can answer what, but cannot deeply interpret where or where the difference is.

[0042] The cause of the defect is that the network structure lacks an explicit and active multi-scale feature fusion module. The features of different levels extracted in the encoding process, such as shallow features C1 / C2 containing details and deep features C5 containing global semantics, fail to effectively interact and weightedly integrate in a unified framework customized for quality evaluation tasks. The feature representation relied on by the model for final decision is suboptimal, and the multi-scale information in the image is not fully utilized.

[0043] 3. The evaluation index is seriously disconnected from human subjective perception, resulting in evaluation results that do not match actual experience.

[0044] Specific defects: the output of the prior art is an objective physical quantity (green rate), while the urban landscape quality is essentially a subjective perception experience of humans (such as beauty, neatness, and pleasantness). There is a huge gap between the physical index and the perceived quality. A street with high green rate but in disarray has high physical index but low perceived quality. The prior art cannot cross this gap.

[0045] The cause of the defect is that the core of its technical route is identification-statistics, and the entire system never takes the landscape quality itself as a direct learning target. The training data of the model is a pixel-level class label, not a quality score label; its loss function aims to optimize classification accuracy, not to approximate human aesthetic judgment. Therefore, the system is essentially a physical element analyzer, not a quality evaluator.

[0046] The above defects make the prior art unable to accurately and comprehensively evaluate the urban landscape quality, resulting in insufficient urban landscape quality evaluation accuracy.

[0047] The embodiment of the present application provides a kind of urban landscape quality evaluation method and system fusing street view image multi-scale feature, to solve the technical problems that the prior art has inherent defects such as single evaluation dimension, shallow feature utilization and index and perception disconnection, can only realize primary element evaluation, cannot accurately and comprehensively evaluate the urban landscape quality, resulting in insufficient urban landscape quality evaluation accuracy. With the technical effect of improving the accuracy of urban landscape quality evaluation.

[0048] In order to make the inventive purposes, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the following described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0049] It should be noted that in the embodiments of the present application, the object information and other related data involved in the specific product or technology need to obtain the permission or consent of the object, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the country and region. That is to say, if the data related to the object is involved in the embodiments of the present application, it needs to be obtained with the authorization and consent of the object, the authorization and consent of the relevant department and the compliance with the relevant laws, regulations and standards of the country and region. If personal information is involved in the embodiments, the consent of the individual is required for obtaining all personal information. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0050] It should be noted that the terms "first", "second" and the like in the present application are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation described in the following exemplary embodiments does not represent all implementations consistent with the present disclosure.

[0051] In addition, the term "and / or" in this paper is only a description of the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents an "or" relationship between the front and rear associated objects unless otherwise specified.

[0052] Embodiment one: Please refer to Figure 2 The present application provides a cityscape quality evaluation method fusing street view image multi-scale features, comprising: step 101, in response to an evaluation request, obtaining street view images and geographic position information of each sampling point in the evaluation area.

[0053] In response to the request, it means receiving the instruction or trigger signal initiated by the user or the city planning management system for the quality evaluation of the cityscape of a specific geographic range.

[0054] The to-be-evaluated region refers to a geographical space range specified in the instruction or signal and required to be evaluated in terms of the urban landscape quality, such as an administrative block, a planned new area, or a specific road and its surrounding area.

[0055] Each sampling point refers to a series of discrete geographical positions determined in the to-be-evaluated region according to a preset spatial distribution rule, each sampling point representing an independent evaluation unit for collecting and analyzing the visual environment of the surrounding streets, having a unique geographical coordinate, and usually corresponding to a group of street view images obtained from different horizontal directions of the point.

[0056] The street view image refers to a digital image collected by a panoramic camera mounted on a mobile vehicle or a fixed facility, which can reflect the building facade, greenery, public facilities, sky, and ground conditions on both sides of the street in a horizontal 360-degree view.

[0057] The geographical position information refers to the spatial position data accurately corresponding to each street view image, specifically the longitude and latitude coordinates of the image collection point in the Earth coordinate system.

[0058] In the embodiment of the present application, the instruction or trigger signal for urban landscape quality evaluation is received, and then the street view images of the to-be-evaluated region and the longitude and latitude coordinates of the image collection points are obtained.

[0059] Step 102, pre-processing the street view images to obtain pre-processed street view images.

[0060] In the embodiment of the present application, a series of pre-processing operations are performed on the street view images to obtain pre-processed street view images.

[0061] Step 103, inputting the pre-processed street view images into a pre-trained landscape quality evaluation model to obtain the landscape quality score value of each sampling point, wherein the pre-trained landscape quality evaluation model is used to extract visual features of different scales from the pre-processed street view images and fuse them, and output the landscape quality score value based on the fused visual features.

[0062] The landscape quality evaluation model refers to a specially trained deep learning neural network architecture running in an end-to-end manner.

[0063] The landscape quality score value refers to a continuous numerical result output by the above-mentioned landscape quality evaluation model after calculating a single input street view image, which is a scalar value and can be set in the range of 1 to 5, used to quantitatively represent the level of the street scene photographed by the image in the comprehensive dimension of urban landscape quality.

[0064] It should be noted that the specific range of the appearance quality score value can be adjusted according to the training data or the scoring standard, and is not specifically limited here.

[0065] It is worth mentioning that the appearance quality score value is the comprehensive output of the model after comprehensively perceiving and judging the multi-dimensional elements (such as greening, building, space, neatness, etc.) and their mutual coordination relationship in the image, which is used to approximate the subjective overall evaluation of the scene by human experts. This value directly reflects the level of perceived quality such as aesthetic, amenity and coordination of the measured place in vision.

[0066] It should be noted that the training process of the appearance quality evaluation model is as follows: a large general image dataset is constructed to generate appearance quality score labels, specifically: A review team composed of city planners, architects, and landscape designers (not less than 5 people) is formed to independently score a large number of preprocessed street view images (such as using a questionnaire). The scoring standard covers multiple dimensions such as greening abundance, building coordination, space neatness, and visual aesthetics, and finally requires experts to give an overall appearance quality score (1-5) based on multi-dimensional perception. For each image, the arithmetic mean of all expert scores is calculated as its target score label Y.

[0067] During model training, the Smooth L1 Loss is used as the loss function, and the formula is as follows: In the formula, Y hat is the predicted value, Y is the true value, and |Y hat -Y| is the absolute difference between the predicted value and the true value.

[0068] At the same time, the Adam optimizer is used for parameter optimization, the initial learning rate is set to 1e-4, the batch training strategy (Batch Size is usually 16 or 32) is adopted, and the learning rate decay strategy is used.

[0069] The model is repeatedly iterated (such as 100 cycles) on a large general image dataset composed of training set, validation set and test set (proportion is 7:2:1) until the performance on the validation set tends to be stable. Finally, the model parameters with the best performance on the validation set are saved, that is, the pre-trained appearance quality evaluation model is obtained. The data in the training set, validation set and test set are obtained through the street map service interface, and the image is processed by screening, size scaling and pixel value standardization. The specific steps are similar to steps 202-204, and are not repeated here.

[0070] During the training process, a learning rate decay strategy is adopted, such as halving the learning rate when the validation set loss does not decrease for 5 consecutive periods, which can automatically reduce the step size when the model approaches the local optimum, avoid oscillation around the optimal value caused by excessive learning rate, make the parameter update more finely approach the optimal solution, and at the same time, through dynamic adjustment of the learning rate, help the model find a better convergence direction after the loss platform period, prevent falling into a local minimum value, and inhibit the risk of overfitting in the later period, and improve the generalization ability of the model.

[0071] It should be noted that the embodiment is implemented under the Python 3.8 and PyTorch 1.9.0 framework, and the NVIDIA GeForce RTX 3090 GPU is used for model training.

[0072] It is worth mentioning that the urban style quality evaluation model of the embodiment has excellent performance on the test set, and the Pearson correlation coefficient between the predicted score and the expert overall score reaches 0.89, which is significantly higher than the correlation coefficient of 0.62 obtained by using only the single indicator of green view rate for evaluation in the prior art, and the root mean square error (RMSE) of the model on the test set is significantly reduced from 0.51 to 0.38, with a performance improvement of about 25%. This data comparison fully proves that the internal representation of the model can better distinguish between hierarchical landscapes and monotonous greenery, as well as coordinated building facades and chaotic building appearances, and effectively overcome the defect of single evaluation dimension in the prior art, and the output score can more comprehensively and accurately reflect the comprehensive quality of urban style, greatly improving the urban style quality evaluation precision.

[0073] In the embodiment of the application, the pre-processed street view image is input into the pre-trained style quality evaluation model to obtain a series of style quality score values, so that the evaluation result is more comprehensive and more consistent with human comprehensive perception of urban space, and more accurately reflects the comprehensive quality.

[0074] Step 104, based on the geographical position information and the corresponding style quality score value of each sampling point, the urban style quality evaluation result of the region to be evaluated is obtained.

[0075] The urban style quality evaluation result refers to the graphical or dataized result that can comprehensively reflect the spatial distribution characteristics of the urban style quality in the region to be evaluated, which is generated by spatial analysis technology based on the style quality scores and geographical position information of all sampling points.

[0076] In the embodiment of the application, the score and the corresponding geographical position information of the sampling point are fused to obtain an urban style quality heat map, realizing the sublimation from single-point image scoring to regional spatial cognition, and converting the abstract numerical score into a graphical result that can be directly used for urban planning, management decision and public understanding.

[0077] Embodiment two: Referring to Figure 3 The application provides another urban landscape quality evaluation method fusing street view images and multi-scale features, comprising the following steps: 201, in response to an evaluation request, acquiring street view images and geographic position information of each sampling point in a region to be evaluated.

[0078] Preferably, step 201 can comprise the following sub-steps: S11, based on a preset road network sampling density, arranging sampling points on a road network in the region to be evaluated.

[0079] The road network sampling density refers to a numerical parameter preset for controlling the spatial distribution interval of the sampling points, for example, one sampling point every 50 meters, which can be set according to actual needs and will not be described here.

[0080] The road network refers to a collection of center line vector data of all streets and roads through which vehicles or pedestrians can pass in the region to be evaluated.

[0081] It can be understood that the vector road network data of the region to be evaluated is read, and then based on the preset road network sampling density, a series of coordinate points are automatically generated along the center line of each road at a specified interval by using the linear referencing and point generation function in the geographic information system (GIS) tool, which are the sampling points used for collecting street view images in the subsequent steps.

[0082] In the embodiment of the application, based on the preset road network sampling density, a series of coordinate points are automatically generated along the center line of each road at a specified interval by using the geographic information system, ensuring that the sampling point systemically and uniformly covers the passable street space of the entire evaluation region.

[0083] S12, determining the longitude and latitude coordinates of each sampling point as the geographic position information according to the arranged sampling points.

[0084] The longitude and latitude coordinates refer to a pair of numerical values used to uniquely identify any position on the surface of the earth represented by a geographic coordinate system, wherein the longitude represents the east-west direction position and the latitude represents the south-north direction position.

[0085] In the embodiment of the application, the pair of longitude and latitude values automatically generated in the arrangement process and corresponding to each sampling point are directly read or exported as the geographic position information of the sampling point.

[0086] S13, calling a street view map service interface, acquiring the street view images of each sampling point based on the geographic position information of each sampling point and a preset collection direction.

[0087] The street view map service interface refers to an application programming interface provided by an Internet map service provider (such as Baidu Map, Tencent Map, etc.) to allow developers to request and obtain street view image data of a specified location through programming.

[0088] The preset collection direction refers to a preset orientation parameter for obtaining different view angle images from a single sampling point, which usually includes four basic horizontal directions of east, south, west, and north.

[0089] It can be understood that for the latitude and longitude coordinates of each sampling point, the street view images of multiple view angles of all sampling points are obtained in batches by sending an HTTP / HTTPS request to the corresponding street view map service interface according to the preset collection direction, the request carrying the latitude and longitude coordinates of the point and the target direction angle and other parameters, the interface responding to the request and returning the street view image data of the corresponding position and view angle, receiving and saving the images, and structurally associating and storing them with the corresponding sampling point coordinates, collection direction and other information.

[0090] It is worth mentioning that the street view image data is usually a picture file in JPEG or PNG format.

[0091] In the embodiment of the application, all the multi-view street view images of the sampling points are collected by calling the street view map service interface according to the latitude and longitude coordinates of each sampling point and the preset collection direction.

[0092] Step 202, determine whether the street view image meets the preset quality standard, and if not, perform rejection processing to obtain an effective street view image.

[0093] The preset quality standard refers to a series of quantitative or qualitative indicators that are preset for measuring whether a single street view image is clear and usable, and whether it is suitable for inputting into a model for evaluation, including but not limited to: overall image clarity (no serious motion blur), proper exposure (no overexposure or underexposure), no obvious large-area occlusion (such as large vehicles temporarily parked, construction fences, dense pedestrian occlusion, etc.).

[0094] The effective street view image refers to an image that is determined to meet the preset quality standard, with clear and complete image content, and no obvious noise or occlusion interference in the quality screening process.

[0095] If a street view image does not meet the requirements in any of the preset quality indicators, for example, there is serious motion blur, overexposure / underexposure, or is significantly occluded by a large area of temporary occlusion, the image will be determined to not meet the preset quality standard.

[0096] It is worth mentioning that the evaluation can be combined with automatic scripts and manual verification. For each image, the quality score or output quality classification result is calculated. If the quality evaluation result of the image does not meet the preset quality threshold or judgment rule, the image is marked as a low-quality image and automatically excluded. Finally, all images that pass the quality check are retained to form an effective street view image set.

[0097] It should be noted that the above steps of judging the quality of the street view image can use various mature image quality evaluation methods to realize the above judgment, for example: 1. Sharpness judgment: Gradient function (such as Laplacian operator, Tenengrad function), frequency domain analysis, etc. can be used to calculate the image sharpness score, and a threshold is set to judge whether there is serious motion blur; 2. Exposure judgment: The histogram distribution of the RGB channels of the image can be analyzed to calculate the overall brightness, contrast, or overexposure / underexposure pixel ratio of the image, and compared with the preset threshold to judge whether the exposure is appropriate; 3. Obstruction judgment: A general target detection model (such as YOLO, Faster R-CNN) or semantic segmentation model based on deep learning can be used to identify whether there is a large area of temporary obstruction in the image, and the pixel ratio or visual saliency is used for judgment.

[0098] In this embodiment, the specific quality judgment algorithm is not limited, and any existing image processing or computer vision method that can achieve the above quality screening target can be applied to this step.

[0099] In the embodiment of the application, it is judged whether the preprocessed street view image meets the overall sharpness, appropriate exposure, and no obvious large-area obstruction. If it meets the requirements, it is regarded as an effective street view image. If it does not meet the requirements, it is excluded to avoid low-quality noise data interfering with the model judgment and improve the accuracy and robustness of the overall evaluation.

[0100] Step 203, scaling the effective street view image according to a preset size to obtain an intermediate street view image.

[0101] The preset size refers to a pre-set, unified target image size, for example, 512 pixels x 512 pixels.

[0102] The intermediate street view image refers to an image whose spatial size has been unified to a standard preset size after scaling.

[0103] It can be understood that all images, regardless of their original resolution, are scaled to the preset size by bilinear interpolation or similar algorithms.

[0104] In the embodiment of the present application, the effective street view image is uniformly scaled to 512 pixels x 512 pixels, ensuring that each batch of input data received by the model has a consistent width and height, meeting the underlying requirements of the deep learning framework for batch data tensor shape consistency, ensuring the stability and efficiency of the forward propagation and reverse gradient calculation of the model, and fundamentally improving the processing efficiency, training stability and consistency of the final inference result of the model.

[0105] Step 204, performing pixel value standardization processing on the intermediate street view image to obtain a preprocessed street view image.

[0106] It can be understood that the RGB three-channel pixel values of the intermediate street view image are first converted from the integer range of 0-255 to the floating point number range of 0.0-1.0 (i.e. each channel pixel value is divided by 255.0), and then the channel mean (e.g. [0.485, 0.456, 0.406]) and standard deviation (e.g. [0.229, 0.224, 0.225]) precomputed on a large general image dataset are further used to perform standardization calculation on the pixel values of each channel, i.e. standardized pixel value = (pixel value / 255.0-mean) / standard deviation, so that the pixel value distribution of all input images is adjusted to the data distribution range adapted and expected by the model backbone network in the pre-training stage.

[0107] In the embodiment of the present application, the intermediate street view image is first normalized, and then standardized, so that the preprocessed street view image meets the requirements of spatial size uniformity and pixel value distribution normalization.

[0108] Step 205, inputting the preprocessed street view image into the pre-trained style quality evaluation model to obtain the style quality score value of each sampling point, wherein the pre-trained style quality evaluation model is used to extract visual features of different scales from the preprocessed street view image and fuse the visual features, and output the style quality score value based on the fused visual features.

[0109] Preferably, step 205 can include the following substeps: The pre-trained style quality evaluation model includes an encoder, a multi-scale feature fusion module and a regressor.

[0110] S21, inputting the preprocessed street view image into the encoder to perform multi-level feature extraction on the preprocessed street view image through the encoder to obtain visual feature maps of different scales.

[0111] The encoder refers to a deep convolutional neural network pre-trained on a large image dataset, which can specifically use ResNet-50 or ResNet-101 as the feature extraction backbone network, and the original fully connected classification layer has been removed.

[0112] Different scale visual feature maps refer to a set of feature images outputted by the encoder from different depths (levels) of convolutional layers, which differ in spatial resolution (size) and semantic abstraction level.

[0113] It can be understood that when the image is input into the encoder, the network performs forward propagation through the stacked convolutional layers, pooling layers and the like, and in this process, the encoder outputs feature maps from different depths (i.e. different levels) of convolutional layers.

[0114] For example, the outputs of four representative convolutional groups (such as Conv2, Conv3, Conv4, Conv5) in the encoder are extracted and denoted as C2, C3, C4 and C5 respectively. These feature maps have different spatial sizes and semantic abstraction levels, that is, the spatial sizes decrease successively, but the semantic abstraction levels increase successively. Among them, the shallow feature maps have high spatial resolution and contain rich edge, texture and other detailed information, and the deep feature maps have low spatial resolution but contain higher-level and more global semantic information (such as abstract representation of concepts such as buildings and vegetation).

[0115] In the embodiment of the present application, the encoder can automatically and efficiently extract multi-level visual representations from standardized images by using its deep convolutional structure, from shallow to deep and from fine to coarse. The shallow feature maps capture the edges, textures and local details in the image, providing a microscopic basis for the evaluation of the style, and the deep feature maps integrate more context information to form a high-level semantic understanding of the scene elements and their layout relationships. This way, the model can grasp both the local fine features and the global structural information of the urban style, enhancing the model's ability to analyze complex urban style scenes.

[0116] S22, input different scale visual feature maps into a multi-scale feature fusion module, and perform fusion processing on the visual feature maps through the multi-scale feature fusion module to obtain target visual feature maps.

[0117] Preferably, S22 can include the following sub-steps: The different scale visual feature maps include shallow feature maps and deep feature maps.

[0118] The multi-scale feature fusion module includes a channel adjustment unit, a cascaded upsampling and addition unit, and a convolutional neural network block.

[0119] S22a, adjust the number of channels of the shallow feature maps and the deep feature maps to obtain adjusted shallow feature maps and deep feature maps.

[0120] Understandably, a 1x1 convolutional layer is applied to each feature map using a channel adjustment unit. The main function of this 1x1 convolutional layer is to perform linear transformation and dimensionality reduction (or dimensionality increase, depending on the situation) of the channel dimension, and to uniformly adjust the number of channels to a preset value, such as 256. This transforms the feature maps with different numbers of channels into adjusted shallow feature maps and adjusted deep feature maps with the same channel depth.

[0121] Continuing with the previous example, we use 1x1 convolutional layers to uniformly adjust the number of channels in C2, C3, C4, and C5 to 256 to reduce computational complexity and unify the dimensions, resulting in {F2, F3, F4, F5}.

[0122] In this embodiment of the invention, 1x1 convolution is used to achieve preliminary recombination and compression of shallow and deep feature maps, which helps to reduce the amount of subsequent computation and extract more compact feature representations.

[0123] S22b. The adjusted deep feature map is upsampled sequentially to the size of the adjacent shallow feature map, and the features are added to the corresponding shallow feature map to obtain a multi-level fused feature map.

[0124] Understandably, by adopting a feature fusion direction from deep (top) to shallow (bottom), and through cascaded upsampling and addition units, the deep feature map, which has the richest semantic information but the smallest spatial size, is first upsampled to the same size as the adjacent shallower feature map with a larger spatial size using bilinear interpolation. Then, the upsampled deep feature map is added element-wise to the shallower feature map to obtain the first-level fused feature map. Next, the first-level fused feature map is upsampled again to the size of the next adjacent shallower feature map with a larger spatial size, and added element-wise to the shallower feature map to obtain the second-level fused feature map. Finally, the second-level fused feature map is upsampled to the size of the shallowest feature map with the largest spatial size, and added element-wise to the shallowest feature map to obtain the third-level fused feature map.

[0125] Continuing from the previous example, upsample F5 to the same size as F4 and add them together to get M4; upsample M4 to the same size as F3 and add them together to get M3; upsample M3 to the same size as F2 and add them together to get M2.

[0126] In this embodiment of the invention, the adjusted deep feature map is sequentially upsampled to the size of the adjacent shallow feature map, and the features are added with the corresponding shallow feature map to obtain a multi-level fused feature map. The global semantic information contained in the deep features is gradually injected into and enriched into the adjacent shallow features with more precise spatial details, thereby realizing the gradual fusion and complementarity of semantic information and spatial details.

[0127] S22c, inputting the multi-level fused feature map into a preset convolutional neural network block for feature enhancement processing to obtain a target visual feature map.

[0128] The preset convolutional neural network block refers to a light-weight module sequentially stacked by multiple standard layers, which can sequentially include a 3x3 convolutional layer, a batch normalization layer and a ReLU activation function layer.

[0129] It is worth mentioning that the 3x3 convolutional layer is responsible for further local perception and feature transformation of the fused features, the batch normalization layer is used to stabilize the feature distribution, accelerate the convergence of potential subsequent training and has a certain regularization effect, and the ReLU activation function introduces nonlinearity to enhance the expression ability of the features.

[0130] In the embodiment of the application, the multi-level fused feature map is input into the convolutional neural network block for feature enhancement processing, which can smooth the noise or discord introduced in the fusion process, further refine and integrate information from different scales, and strengthen the visual patterns beneficial to the style quality evaluation task. The finally output feature map is the target visual feature map.

[0131] S23, inputting the target visual feature map into a regressor, and performing mapping and regression calculation on the target visual feature map by the regressor to obtain a style quality score value.

[0132] Preferably, S23 can include the following sub-steps: S23a, performing global average pooling processing on the target visual feature map to obtain a global feature vector.

[0133] The global feature vector refers to a one-dimensional numerical vector obtained by performing global average pooling processing on the target visual feature map.

[0134] It can be understood that the target visual feature map (a three-dimensional tensor with height H, width W and channel number C) calculates the average value of all pixel values in each channel along its spatial dimension (i.e. H and W), and then condenses the spatial distribution information of each channel in the original feature map into a single scalar value. Finally, the average values of all C channels are sequentially spliced to form a one-dimensional vector with a length of C.

[0135] It should be noted that in the embodiment, the length of C is 256.

[0136] In the embodiment of the application, the global average pooling processing is performed on the target visual feature map to obtain a global feature vector with a length of 256, which aggregates the global information of the entire image in all feature dimensions, retains the feature semantics represented by the channel dimension, and completely eliminates the spatial position information of the features, thereby enhancing the robustness of the features.

[0137] S23b, input the global feature vector into a multi-layer fully connected network, and perform dimension reduction and nonlinear transformation on the global feature vector through the multi-layer fully connected network to obtain a style quality score value.

[0138] The multi-layer fully connected network refers to a neural network structure sequentially connected by two or more fully connected layers, and is used to map the input high-dimensional feature vector to a target output space.

[0139] It can be understood that the global feature vector is input into the network, the first layer performs feature compression and nonlinear activation, extracts abstract perception factors highly related to the style quality, and the second layer comprehensively weights these factors to directly regress a continuous scalar, which is the style quality score value, and the numerical range (such as 1-5 points).

[0140] For example, the multi-layer fully connected network (or called regression head) is sequentially composed of two fully connected layers. The first fully connected layer receives the global feature vector, maps its dimension to a lower intermediate dimension (such as 256 to 64) through a weight matrix, and applies an activation function such as ReLU to introduce nonlinearity after linear transformation. The second fully connected layer further maps the intermediate dimension feature to the final output dimension 1 (such as 64 to 1), and this layer usually does not set an activation function to output a continuous real value.

[0141] In the embodiment of the application, the global feature vector is gradually reduced in dimension and nonlinearly transformed through the multi-layer fully connected network, so that the model can robustly map high-dimensional and abstract global image features to a single quality score that is easy to understand and compare and reflects human comprehensive perception.

[0142] Step 206, based on the geographical position information of each sampling point and the corresponding style quality score value, obtaining the urban style quality evaluation result of the region to be evaluated.

[0143] Preferably, step S206 can include the following sub-steps: S31, based on the geographical position information of each sampling point and the corresponding style quality score value, generating a continuous distribution map of style quality covering the region to be evaluated through a spatial interpolation algorithm, as the urban style quality evaluation result of the region to be evaluated.

[0144] It can be understood that the latitude and longitude coordinates of each sampling point are associated with the corresponding landscape quality score value to form a set of point-like evaluation data with spatial attributes, and then the spatial interpolation analysis function of the geographic information system is used, for example, the Kriging interpolation method or the inverse distance weighting (IDW) interpolation method, to calculate and generate a continuous distribution of landscape quality score surface covering the entire evaluation area based on the score data of the discrete points, and the continuous score surface is visualized and rendered to generate an intuitive urban landscape quality evaluation result, and the typical form is a city landscape quality heat map.

[0145] It is worth mentioning that in the city landscape quality heat map, different colors intuitively represent the landscape quality scores of different geographical locations, thereby clearly showing the spatial distribution pattern of the landscape quality in the region, the core advantage area and the short board area that needs to be improved, wherein: the high score area is usually represented by dark green or blue, corresponding to the area with high score value (for example, score 4.0-5.0), which can intuitively indicate those streets or blocks with coordinated buildings, rich greenery, clean and orderly space, and good overall visual experience; the medium score area is represented by yellow or light green transition, corresponding to the medium score (for example, 2.5-4.0), which can reflect the area with general landscape quality; the low score area is highlighted by orange or red, corresponding to the area with low score (for example, 1.0-2.5), which can point to specific sites with disordered building facades, insufficient greenery, occupied public space or visual discord.

[0146] For example, in a city landscape quality heat map, the historical commercial center of the city may appear as a coherent green patch, indicating that the overall landscape quality is good; while a newly built but disordered planning logistics park or an old industrial area to be updated may be displayed as a prominent red area in the map, directly revealing its short board position in the city landscape.

[0147] In the embodiment of the present application, the latitude and longitude information of the sampling points and the landscape quality score value are fused to obtain the city landscape quality heat map, and through this visualization method, the decision maker can instantly grasp the overall landscape pattern and accurately locate the specific spatial position that needs to be improved.

[0148] It is worth mentioning that in the practical application of the embodiments of the present application, the new district of a city is evaluated, and a logistics park with high green view rate but low landscape quality (green view rate 35%, evaluation score of the present application 1.8 / 5.0) is successfully identified. The prior art will misjudge it as a high-quality area due to its high greening rate. Compared with the prior art, the evaluation accuracy of the urban landscape quality of the present application is higher. In addition, it usually takes 2-3 weeks to form a team to conduct on-site research on a 10 square kilometer area, and the cost is about 200-300 thousand yuan. The present application only needs to pay a small amount of cloud computing resource fee, and can complete the same work in 1 day, with an efficiency improvement of more than 20 times and a cost reduction of 1 / 10 or even lower.

[0149] It is worth mentioning that the embodiments of the present application give scores to different streets through intelligent identification of assignment, and can quickly determine the street blocks that need to be improved according to the score weight, such as building facade sequence management (preventing illegal construction), building front area external arrangement order management (preventing random arrangement), slow lane unobstructed management (avoiding random parking and obstruction), and motor vehicle lane facility sequence (railing skew), which further improves the urban management efficiency.

[0150] Embodiment three: Please refer to Figure 4 The present application provides a kind of urban landscape quality evaluation system of fusion street view image multiscale feature, comprising: data acquisition module 101, for responding to evaluation request, the street view image and geographic position information of each sampling point in the region to be evaluated are acquired; Image preprocessing module 102 is used to pre-process the street view image to obtain the pre-processed street view image; Quality score module 103 is used to input the pre-processed street view image into the pre-trained landscape quality evaluation model to obtain the landscape quality score value of each sampling point, wherein the pre-trained landscape quality evaluation model is used to extract different scale visual features from the pre-processed street view image for fusion, and output the landscape quality score value based on the fused visual features; Evaluation result generation module 104 is used to obtain the urban landscape quality evaluation result of the region to be evaluated based on the geographic position information of each sampling point and the corresponding landscape quality score value.

[0151] Since the above is a kind of urban landscape quality evaluation system of fusion street view image multiscale feature, its implementation principle is consistent with the method of fusion street view image multiscale feature, for the convenience and brevity of description, those skilled in the art can clearly understand the specific working process of the above-described system and module, which can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

[0152] Embodiment four: The electronic device of the embodiment of the application comprises a memory and a processor, and the memory stores a computer program; when the computer program is executed by the processor, the processor executes the cityscape quality evaluation method of fusing street view image multi-scale features according to any one of the above embodiments.

[0153] The memory can be an electronic memory such as a flash memory, an EEPROM (Electrically Erasable Programmable Read-Only Memory), an EPROM, a hard disk or a ROM. The memory has a storage space for program codes for executing any of the method steps described above. For example, the storage space for program codes can comprise individual program codes for implementing respective steps in the above method. These program codes can be read from or written to one or more computer program products. These computer program products comprise program code carriers such as hard disks, compact disks (CDs), memory cards or floppy disks. The program codes can be compressed in a suitable form, for example. These codes, when executed by a computing processing device, cause the computing processing device to perform the respective steps in the method described above.

[0154] Embodiment five: The embodiment of the application provides a computer readable storage medium, which stores a computer program, and the computer program is executed to implement the cityscape quality evaluation method of fusing street view image multi-scale features according to any one of the above embodiments.

[0155] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0156] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the above-described device embodiments are merely schematic; the division of the units is merely a logical function division; an actual implementation can be another division manner, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0157] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0158] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0159] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in part, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods according to the embodiments of the present application. The aforementioned storage medium includes: various memories (such as a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a disk, or an optical disc, and the like) that can store program codes.

[0160] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for evaluating urban landscape quality by integrating multi-scale features of street view images, characterized in that, include: In response to the evaluation request, obtain street view images and geographic location information of each sampling point in the area to be evaluated; The street view image is preprocessed to obtain a preprocessed street view image; The preprocessed street view image is input into a pre-trained landscape quality evaluation model to obtain the landscape quality score value of each sampling point. The pre-trained landscape quality evaluation model is used to extract visual features of different scales from the preprocessed street view image, fuse them, and output the landscape quality score value based on the fused visual features. Based on the geographical location information of each sampling point and the corresponding landscape quality score, the urban landscape quality evaluation result of the area to be evaluated is obtained.

2. The urban landscape quality evaluation method based on the fusion of multi-scale features of street view images according to claim 1, characterized in that, The acquisition of street view images and geographic location information of each sampling point within the area to be evaluated includes: Based on the preset road network sampling density, sampling points are set up on the road network of the area to be evaluated; Based on the deployed sampling points, the latitude and longitude coordinates of each sampling point are determined as the geographical location information; The street view map service interface is invoked to obtain street view images of each sampling point based on the geographical location information of each sampling point and the preset collection direction.

3. The urban landscape quality evaluation method based on the fusion of multi-scale features of street view images according to claim 1, characterized in that, The preprocessing of the street view image to obtain a preprocessed street view image includes: Determine whether the street view image meets the preset quality standard. If it does not meet the standard, it is discarded to obtain a valid street view image. The effective street view image is scaled up to a preset size to obtain an intermediate street view image; The intermediate street view image is subjected to pixel value normalization processing to obtain the preprocessed street view image.

4. The urban landscape quality evaluation method based on the fusion of multi-scale features of street view images according to claim 1, characterized in that, The pre-trained landscape quality evaluation model includes an encoder, a multi-scale feature fusion module, and a regressor. The step of inputting the preprocessed street view image into a pre-trained landscape quality evaluation model to obtain the landscape quality score value of each sampling point includes: The preprocessed street view image is input into the encoder, and the encoder performs multi-level feature extraction on the preprocessed street view image to obtain visual feature maps at different scales. The visual feature maps at different scales are input into the multi-scale feature fusion module, and the multi-scale feature fusion module performs fusion processing on the visual feature maps to obtain the target visual feature map. The target visual feature map is input into the regressor, and the regressor performs mapping and regression calculations on the target visual feature map to obtain the appearance quality score.

5. The urban landscape quality evaluation method based on the fusion of multi-scale features of street view images according to claim 4, characterized in that, The visual feature maps at different scales include shallow feature maps and deep feature maps; The step of inputting the visual feature maps of different scales into the multi-scale feature fusion module, and fusing the visual feature maps through the multi-scale feature fusion module to obtain the target visual feature map includes: The number of channels in the shallow feature map and the deep feature map is adjusted to obtain the adjusted shallow feature map and the deep feature map; The adjusted deep feature map is sequentially upsampled to the size of the adjacent shallow feature map, and the features are added to the corresponding shallow feature map to obtain a multi-level fused feature map; The multi-level fused feature map is input into a preset convolutional neural network block for feature enhancement processing to obtain the target visual feature map.

6. The urban landscape quality evaluation method based on the fusion of multi-scale features of street view images according to claim 4, characterized in that, The step of inputting the target visual feature map into the regressor, and using the regressor to map and regress the target visual feature map to obtain the appearance quality score includes: The target visual feature map is subjected to global average pooling to obtain a global feature vector; The global feature vector is input into a multilayer fully connected network, and the multilayer fully connected network performs dimensionality reduction and nonlinear transformation on the global feature vector to obtain the landscape quality score.

7. The urban landscape quality evaluation method based on the fusion of multi-scale features of street view images according to claim 1, characterized in that, The evaluation result of the urban landscape quality of the area to be evaluated is obtained based on the geographical location information of each sampling point and the corresponding landscape quality score, including: Based on the geographical location information of each sampling point and the corresponding landscape quality score, a continuous distribution map of landscape quality covering the area to be evaluated is generated through a spatial interpolation algorithm, which serves as the evaluation result of the urban landscape quality of the area to be evaluated.

8. A system for evaluating urban landscape quality by integrating multi-scale features of street view images, characterized in that, include: The data acquisition module is used to respond to evaluation requests and acquire street view images and geographic location information of each sampling point in the area to be evaluated; An image preprocessing module is used to preprocess the street view image to obtain a preprocessed street view image; The quality scoring module is used to input the preprocessed street view image into a pre-trained landscape quality evaluation model to obtain the landscape quality score value of each sampling point. The pre-trained landscape quality evaluation model is used to extract visual features of different scales from the preprocessed street view image, fuse them, and output the landscape quality score value based on the fused visual features. The evaluation result generation module is used to obtain the urban landscape quality evaluation result of the area to be evaluated based on the geographical location information of each sampling point and the corresponding landscape quality score.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 7, which is a method for evaluating urban landscape quality by fusing multi-scale features of street view images.

10. A computer-readable storage medium, characterized in that, The computer program stores a method for evaluating urban landscape quality by fusing multi-scale features of street view images as described in any one of claims 1 to 7, which can be loaded by a processor.

Images